AI detectors and the fairness gap: read the evidence before you trust the score

This blog was kindly authored by Anh Do, Founder and CEO of MAAS EdTech.

Ask an AI detector whether a student used a chatbot and it will give you a confident answer. Ask the researcher whether the detector is right and that confidence falls away. When seven widely used detectors were tested on 91 essays written for TOEFL (the standard English-proficiency exam) by people who do not speak English as a first language, they wrongly flagged an average of 61 per cent of that genuine human writing as AI-generated. On essays by US pupils writing in their first language, the same tools were far more accurate.

That gap should shape how universities use these tools because it shows the errors are not random. Many detectors work by measuring how predictable a piece of writing is: varied, surprising phrasing looks human; while plain and repetitive phrasing looks machine-made. Someone writing carefully in a second language tends to use a narrower range of words and sentence patterns, so the very habit of the diligent multilingual student is the habit the tool reads as a robot. The people least able to absorb a misconduct accusation are the ones the tools are most likely to accuse.

When Turnitin launched its detector, it reported a false-positive rate of about one per cent, meaning it would wrongly flag roughly one in a hundred genuine papers. Vanderbilt University did the arithmetic on its own figures: it had run around 75,000 papers through Turnitin the previous year, so one per cent implied about 750 pieces of real student work wrongly branded as AI. It judged the tool too unreliable for assessment and switched it off. Several other US universities stepped back from automated detection in the same period. A number that sounds trivial in a vendor’s brochure becomes hundreds of individuals the moment it meets a real cohort, and each of them is then asked to prove they did their own work.

None of this means detectors are always wrong, and the argument is stronger for admitting where they are not. One study of 160 genuine student assignments found no false positives at all, with every human paper scored at zero AI likelihood. Vendors have published rebuttals defending their tools, and broader evaluations find the error rate jumps around from tool to tool rather than being uniformly high. Put together, the honest reading is not that detection never works but that there is no universal accuracy figure: the false-positive rate swings widely with the tool, the sample and the settings, and it is worst precisely on the writing of students who learned English later.

This is not only an American story. HEPI’s own Student Generative AI Survey 2026 found that 94 per cent of UK undergraduates now use generative AI for assessed work, and the share who put AI-generated text straight into their submissions has risen from three per cent in 2024 to 12 per cent in 2026. As real AI use climbs, so does the pressure to detect it; the same survey records students voicing a clear anxiety about being wrongly accused of misconduct. The UK also teaches one of the world’s largest international student populations, the group these tools misread most. The fairness gap is not an import; it is already sitting in British marking queues.

So what are the learning points for institutions regarding this?

1. A detector score should never be sufficient evidence of misconduct on its own; it can justify a closer look, never a verdict, and the burden of proof should rest with the process, not with the accused student.
2. Any institution using a detector should know and publish its assumed false-positive rate, then read that rate against its own submission volumes and its own share of international students, rather than trusting a headline number.
3. The durable protections are the ones this sector already understands: assessment designed to be hard to fake, and real weight given to the ordinary evidence of a student’s process, their drafts, version history and their ability to talk through their own argument. A fuller review of the evidence is collected here.

Detection tools can have a place in that picture. As a prompt to ask a better question, they are defensible. As the answer, they manufacture unfair outcomes and hand most of them to the students that a fair system should be working hardest to protect.